Manifesto — Execution-Governed AI

AI safety is not what you train. It is what you let execute. The difference is the only thing that matters.

1. I

   Containment is strictly faster than cognition.

   A model that can act faster than its containment cannot be governed by it. Therefore governance must precede execution, not annotate it.

2. II

   Deny-by-default is the only safe default.

   Absence of an explicit ALLOW is not permission. Every meaningful action must affirm itself against identity, capability, policy, ethics, security, consistency, and audit — in that order, every time.

3. III

   Three verdicts, no fourth.

   ALLOW · DENY · SAFE_HALT. A soft fourth state — warn, rewrite, retry, flag — is the seam where containment fails. The kernel will not provide it.

4. IV

   Policy is not in the model.

   Constitutional AI shapes weights. We do not trust weights. Policy lives in a hash-anchored Code Store, loaded by validators at runtime, beyond the reach of the cognition it governs.

5. V

   The audit is the system.

   If a decision is not signed, it did not happen. If a receipt is not in the chain, the action was not governed. Append-only is the only memory the kernel keeps.

6. VI

   Capability is explicit and tiered.

   AC0 through AC5. A capability not granted is a capability denied. There is no honorary class, no implied access, no veteran's exception.

7. VII

   Continuity is a gate.

   STATE_REGISTER fires SAFE_HALT on a missed heart-tick. A kernel that drifts silently is a kernel that does not exist.

8. VIII

   Triumvirate, not majority.

   Galahad attests. Cerberus contains. Codex arbitrates. No two may overrule the third on its own surface. Disagreement is SAFE_HALT, not vote.

9. IX

   Eat the dogfood, in public.

   Every decision this portal makes is published on /witness. Every claim made here is testable on /challenge. The substrate that cannot survive its own scrutiny does not deserve yours.